AZK AI

Language attrition has accelerated since 80% of all digitalization efforts are in English. This is creating a marginalization of huge populaces who have to decide between preserving cultural identity versus marching into a digital future. For these societies, there is an ever-widening gap between English and non-English speakers. Some of these are from Perseo-Arabic-speaking-countries who are 800+ million people.

A major barrier to digitalization where Perseo-Arabic script is used is that there are very few tools to convert handwriting to digital formats as most advanced tools are in English and do not lend to conversion because of  Character set variances and syntactic barriers. Therefore, we have created an AI-based system that takes handwritten documents in native languages and digitizes the native data by understanding its structure.

Our solution, in pilot successfully supports Urdu. However, in subsequent phases, we will be able to extend the system to other Perseo-Arabic variants.

Because Perseo-Arabic languages have unique traits and are cursive in nature, computer recognition of written Perseo-Arabic languages (including Urdu) is more difficult. Typically, conversion of the text for consumption and processing in computers is considered a difficult task compared to Latin scripts and current optical recognition systems often have difficulty in segmentation and OCR of such languages. 

Current Limitations are:

1. Having a Handwritten Document image does not have some basic data processing features like edit, searching, copy, grammar check, etc. Further we always require an agent (Human Interaction) for decision making because we cannot apply rule based approaches on Images.

2. A language such as Arabic or Urdu is extremely challenging to some other languages such as English, French etc. Because these languages are cursive in nature and segmentation in them gets challenging for character extraction. Icing on the cake is their difficult ligature to handle which varies with regions.

This issue is a major barrier to the digital inclusion of all Perseo-Arabic natives speakers. This is faced by all those who use also similar writing systems used in Asia and Africa, such as Arabic, Urdu, Persian (Farsi/Dari), Uyghur, Kurdish, Punjabi, Sindhi, Balti, Balochi, Pashto, Lurish, Urdu, Kashmiri, Rohingya, Somali and Mandinka, among others.

There are a number of issues that hinder the widespread adoption of Perseo-Arabic languages like arabic, urdu due to it bidirectional language in nature. Additionally, the character shape is context sensitive. Therfore mostly these languages data is not easy to manipulate. Nor its character recognition are easy for the traditional OCR systems. 

Our Deep learning based recognition systems have shown encouraging performance even in the presence of challenges mentioned above. The proposed solution will enable deep learning system to learn the variants of handwritten and printed Perseo-Arabic language text and transform them in computer readable content, which can easily be storeable, indexable, and searchable, editable by the native langauge speakers.

Our solution include following simple steps:

1. Upload PDF/Image
2. Send (to server) for Preprocessing
3. Template Matching
4. Identify Handwritten from printed text
5. Spell Checking
6. Store the Data in Digital format as text.

Our solution will include further steps

1. Analyze the digital formated data
2. Take apporiate decsion

The solution affects a large part of the Arabic world where these scripts are mainly used. As mentioned above there are close to 10 major language groups that make up the Perseo-Arabic belt and total 800+ million people. In addition, there are 20 other smaller language groups that use derivative scripts.

For our first test, we have used Urdu as a test case. Urdu is a national language of Pakistan- a population of 220 million. Additionally, it is estimated another 100 million speak or write Urdu.  However, because it does not lend itself to digitisation easily, it is becoming obsolete on the one hand. 

On the other, Urdu language writers continue to use handwritten ledgers for conducting business which obscures transparency and links. examples of these are the land registers, police FIRs, local retailers accounts and so on, which continue to be in handwritten formats because there is no easy method to digitise these records.

To take the next step in this population, it is now imperative that Urdu wordprocessing and recognition is made easier.

We have a user base that we are taking as our initial test case. For them we are digitalizing their systems for faster adoption by their customers.

We are targeting Urdu speakers as our pilot customer base. The government is attempting to transition into using Urdu as a business language from English because there are close to 220 million Urdu speakers whereas only 3 million  English speakers (https://www.thoughtco.com/what...). This is creating a class difference and difference in job availability for the two classes. However, systems are still running into difficulty because of a legacy of English use in the subcontinent. 

Also as native Urdu speakers, it is easier for us to comprehend, develop and test solutions for this market.

Currently, we had developed a protoype that can effectively decode and recognise handwritten Urdu forms. We already attached working proof of our claim as an video in this application. Now we are working on End to End pipeline, which will allow us to start our first pilot in 3-4 months time.

In our attached demo, we have clearly shows how hand written form is converted to digital form (JSON), which can easily store further in databases and it will allow us to index, search and edit this data.

Further we have a plan to sign contract with local government to digitize their land record, and police FIR (first investigation report), which are currently hand written.

We also has identified the multiple businesses which use hand written documents for their daily operations, and currently either they just scanning images or keeping hard copies.

Our solution has the same innovative essence and power like English has done for Eurasian languages by digitalizing the contents and documents.

Arabic and derived scripts such as Urdu and Persian are known as Perso-Arabic scripts are based on alphabet and rules different from those of Latin script. The Perseo-Arabic script is:

• Written in cursive from right to left.

• Have no upper and lower case characters.

• The character shape is context sensitive-  i.e. its shape is dependent on its position within a word - beginning, middle, end of a (sub)word, or in isolation.

• Contains diacritics (dots, hamza, chadda) or small markings that change the sound based on the position relative to its main shape.

• Some combinations of two or three letters have special shapes called “ligatures”. These multiple configurations give a total of more different shapes which is to be compared to the 52 different character shapes of Latin script.

• Arabic and like scripts are segmented into PAWs: “parts of Arabic words”, each consisting of a group of letters. A word is composed of one or more PAWs.

Our solution will help people of Arabic, Persian and Urdu speakers to digitize their documents and content in their native languages without worrying to convert to English. Additionally, 

• Digital Conversion of local dialects becomes easier

• Business can be conducted in local dialect

• It would bring large number of native speakers as part of the digital evolution 

• It will reverse the downfall of local and regional languages

Current 

We are currently piloting and testing our technology. We have fed the system 2000+ documents and the system is now at 81% average accuracy. We are also working with one public organisation automate a part of their handwritten ledger system. At this rate we feel we will be able to roll out a API by 2022 Q2.

1 Year Milestones 

Real impact of the system will begin Q2 2022. In one year we plan to develop an API app for use for other systems and also have an OCR version for mobile-phone. With more data we will be able to train our model to achieve even better accuracy. After Q2 2022, We estimate this will be able to develop a solution to 800k+ documents and impact around 500-600k people. This does not account for the jobs that will now be available for locally trained Urdu medium graduates and the ability to transform organisations due to ability to use Urdu as a medium of instruction.  

This will effectively be the time frame we will be rolling out the system for commercial use. After Q2 2022, we will be able to distribute this API for generalized usage. This can multiply digital inclusion through customers ( of Urdu nationalities) 

Five Year Milestones 

In five years we envisage a change in the local web culture of Pakistan where Urdu can be used for all places where currently there is no alternative but to use English. This includes web services, financial, health and education sectors and many consumer facing organisations like government, taxation, banks,etc where Urdu speaking customers will be able to make informed decisions. 

While we havent an exact number, we envisage a positive culture change in being able to use Urdu as a local language, open more job opportunities for Urdu - medium graduates.

Additionally, since the Urdu application development to commercialisation process is approximately 2 Years. We plan to simultaneously work Q1 2023 to train the application in other local languages such as Baluchi, Sindhi and Arabic which are derivative of the Urdu language - Initially we plan choosing work on two languages at a time, based on demand. This will help us to grow more exponentially than first year and impact between additional served better as well as improve work in many consumer facing organisations like government, taxation, financial inclusion,etc. While we havent an exact number, we envisage a positive culture change in being able to use Urdu as a local language, job  people.

The ultimate success of our system is when banks and other institutions are able to use Urdu as a basic alternate for all work and create dss in their local dialects. this will not only create more job opportunities in data entry, data processing but will also enable the common man and woman and child to comfortably use our system. 

We estimate this will take another 2 years to be in a shape that is adapted by organizations. However, after this system is in place, Urdu medium and English medium schooling will be expected to be at par. At this time, those who are only Urdu speakers are discriminated as all business is either conducted using computers in English or on paper- which lends itself to being archaic. Additionally, only Urdu language speakers are discriminated at many levels. 

We are currently maintaining minimum team to keep running the technology development. Our team includes:

Full Time:

1. Rehan Mushtaq (CEO/CTO - Founder)
2. Arsalan Younus (Lead Data Scientist - CoFounder)
3. Usman Latif (Software Engineer)

Part Time:

1. Rehan Sarwar (Data Annotation Engineer)
2. Ibrahim Saleh (QA Engineer)
3. Muzamil Hassan (Data Annotator)

Rehan Mushtaq:

He has 21 years of experience in Software Development, AI Practioner, Data Science, Team Management, and Technology Consulting. Successfully operate multi millions dollar software services companies.

Native speaker of Urdu, Punjabi and also communicate in Arabic and English. Providing services  in the regions south asia, middle east, and others.

Arslan Yonus:

He has 3+ years experience as a Data Scientist using NLP technologies. He is also native to south asia region. He has expertise and experience of digitalization of handwritten documents.

Usman Latif:

3 Years experience as Full Stack Developer using technologies like Laravel, React, Angular, MySQL and MongoDB.

Our vision is:

1- Build a company that places high value on merit. We hire people based on their abilities not because of their race, gender, language, or any other basis.

2- We respectfully treat all employees as our business partners. We believe the job needs to be done and people need to have systems that do not automatically exclude anyone out. That is the vision that we have for our Pakistan - is to treat everyone equally.

3- We build solutions that help common man and woman to lead respectful, dignified lives and open doors to better livelihoods.

We are applying to Solve because we feel big ideas like ours need big support to turn into reality. Huge segments of society are currently left behind in the digital revolution on the basis of Language. We want it to be easier for the Middle Eastern, South Asian, Central Asian, and African Communities to tap into the power of digital through our impact model. this cannot be done alone- and behind every visible idea are countless entrepreneur supporters, who help push the idea along inches until it succeeds. We envisage a world where language is not a barrier. Therefore we need to connect to many organizations and people. Our top reasons for applying to MIT Solve are:

It gives the opportunity to connect to like-minded people solving similar problems. We feel the global platform the MIT Solve provides helps us develop the global platform needed connections for technical advice, funding opportunities, partnerships for extending our work.

Because our technology used is Deep learning we would like to connect with the MIT experts and faculty working in these areas to advise and verify our direction and mentor us to developing a better solution.

MIT Solve also highlights our problem of digital exclusion for national languages. We feel that raising this issue at a global level will help develop solutions for language attrition and bring all nationalities to par. 

Human Capital: We need to develop a pool of data scientists, consultants, social scientists, and other stakeholders for helping accelerate our journey to market.   

Business model: We are looking for mentorship to build a sustainable model for extending this solution not only for Urdu but for all Perseo Arabic nationals.

Financial: We are looking for mentors and coaches so we can pitch this solution better and obtain more funding for our R&D efforts

Public Relations: This problem needs to be highlighted particularly amongst the development sector. We need champions to preserve our national language(s)

Monitoring & Evaluation: We are looking for mentors to help us in handling the data, deploy correct security measures and help us in refining exact impact measurements

Product / Service Distribution: We are looking for local partners who can take leadership of each language customer base. 

Technology: We are happy to partner with cloud services and MIT faculty and experts to audit our solution and mentor us for building a robust solution.

We would like to partner with Financial and technical partners who can get us to launch faster.

Technical partners: To help us build the best solution to our problem, we would like to partner with organisations like MIT Faculty for mentoring, AWS Cloud services or Online processing and storage, Hardware Companies like Nvidia for faster graphics processing and other such companies.  

Marketing and Channel Development:We would like to accelerate entry into various language markets we need marketing partners in different regions. 

Financial: We are actively looking for investors to develop a pay-as-you-go model for organisations. This is to enable faster adoption.

This program can be adapted to Dari and Pushto. Pakistan is home to 1.44 million Afghani refugees of which 1 million are outside of refugee camps and about half a million in refugee camps. A use case of our product is enabling translations between Arabic Perseo dialects which includes Urdu, Arabic, Dari, and Pushto. This is only possible if a system like us lowers the barriers of word processing in Arabic/variants letters.

Language is one of the major barriers to digital inclusion. We want to ensure that no person is excluded based on language. Therefore we are looking to make sure that this set of languages are digitally processed as well as others. This is the only way we will ensure digital equity of 1+ billion people of the Who have declared Perseo-Arabic language as their national language.

We feel we are qualified to apply for this prize. While natural language processing using deep learning has made great strides in English. We are still a long way to achieving the same in Urdu, Arabic or any other Perseo-Arabic script. Our solution will help National language absorption and reverse attrition.

Our Solutions is based on AI including Deep Learning, NLP, Object Detection and this digital inclusion is currently possible to due to these technologies.

A huge segment in our developing countries is excluded from the digital world due to the language barrier. This solution provides a basic platform that can be used for Arabic, Persian, Urdu communities and their variants. This will then open countless opportunities for learning, jobs, economic development and hope to raise standards of living for the marginalized communities who currently are "digital illiterates".